Modeling scene text features with parametric filter banks and contextual color-shift distribution model

Texts in scene images provide important information for indexing and searching. If these texts can be correctly located, segmented and recognized, they could provide a semantic source for scene understanding. In this paper, we propose a method to model scene text features with structured or semi-structured fonts. The framework is composed by three stages: 1) Tailored filter banks with a parametric parallel detector are used to detect stroke-like areas on multi-scale inputs. These stroke-like areas (candidates), being the potential skeletons of text regions, are then passed through a labeling scheme. 2) A color-shift distribution model is first sampled under a pair of selected variants and then trained in a pool of scene-text images. Candidates are assigned belief values based on this model. 3) A contextual probability function is formulated as a reference to integrate context-free candidates into context-related regions. Modeled features output after a scale fusion process.